Abstract

In this project, we present a rigorous replication of the seminal work on curiosity-driven exploration by self-supervised prediction in reinforcement learning environments. Our study reconstructs the original framework, where an agent's curiosity—quantified as the prediction error of the consequences of its actions—is used as an intrinsic reward mechanism. This approach allows the agent to independently explore and acquire knowledge without external incentives. We replicated experiments in two distinct environments, Breakout and Space Invader, to test the reproducibility of the findings. We followed the procedures for creating an inverse dynamics model that the agent uses to learn visual feature representations of its environment. Our detailed replication contributes to a greater confidence in the robustness of curiosity as an intrinsic motivator for agent exploration.

Document Type

Article

Author's School

McKelvey School of Engineering

Author's Department

Electrical and Systems Engineering

Class Name

Electrical and Systems Engineering Undergraduate Research

Date of Submission

4-24-2024

Share

COinS